Arxiv:2101.05282V3 [Astro-Ph.SR] 2 Feb 2021 at Angular Separations Greater Than About One Arcsecond, Wide Et Al

MNRAS 000,1–26 (2021) Preprint 3 February 2021 Compiled using MNRAS LATEX style ﬁle v3.0

A million binaries from Gaia eDR3: sample selection and validation of Gaia parallax uncertainties

Kareem El-Badry,1,2★ Hans-Walter Rix,2 and Tyler M. Heintz3 1Department of Astronomy and Theoretical Astrophysics Center, University of California Berkeley, Berkeley, CA 94720, USA 2Max-Planck Institute for Astronomy, Königstuhl 17, D-69117 Heidelberg, Germany 3Department of Astronomy, Boston University, 725 Commonwealth Ave., Boston, MA 02215, US

Accepted to MNRAS

ABSTRACT We construct from Gaia eDR3 an extensive catalog of spatially resolved binary stars within ≈ 1 kpc of the Sun, with projected separations ranging from a few au to 1 pc. We estimate the probability that each pair is a chance alignment empirically, using the Gaia catalog itself to calculate the rate of chance alignments as a function of observables. The catalog contains 1.3 (1.1) million binaries with >90% (>99%) probability of being bound, including 16,000 white dwarf – main sequence (WD+MS) binaries and 1,400 WD+WD binaries. We make the full catalog publicly available, as well as the queries and code to produce it. We then use this sample to calibrate the published Gaia DR3 parallax uncertainties, making use of the binary components’ near-identical parallaxes. We show that these uncertainties are generally reliable for faint stars (퐺 & 18), but are underestimated significantly for brighter stars. The underestimates are generally ≤ 30% for isolated sources with well-behaved astrometry, but are larger (up to ∼80%) for apparently well-behaved sources with a companion within . 4 arcsec, and much larger for sources with poor astrometric fits. We provide an empirical fitting function to inflate published 휎휛 values for isolated sources. The public catalog offers wide ranging follow-up opportunities: from calibrating spectroscopic surveys, to precisely constraining ages of field stars, to the masses and the initial-final mass relation of white dwarfs, to dynamically probing the Galactic tidal field. Key words: binaries: visual – stars: evolution – methods: statistical – catalogues – parallaxes

1 INTRODUCTION genuine binaries, in which the two stars have nearly identical proper motions (e.g. Luyten 1971, 1979; Salim & Gould 2003; Chanamé & About half of all solar-type stars are members of binary systems, and Gould 2004; Dhital et al. 2010). Many wide binary searches have a majority of these are so widely separated that the two components specifically targeted high-proper motion stars, which have fewer never interact (e.g. Moe & Di Stefano 2017). With orbital periods phase-space neighbors that can be mistaken for binary companions. ranging from ∼10 to ∼ 108 years, most of these binaries can in If available, parallaxes and radial velocities are also useful for dis- some sense be viewed as clusters of two: the components formed tinguishing binaries from chance alignments (e.g. Close et al. 1990; from the same gas cloud and have orbited one another ever since. Andrews et al. 2017). They thus have essentially the same age, initial composition, and distance, but generally different masses and occasionally different Prior to the Gaia mission (Gaia Collaboration et al. 2016), useful evolutionary phases. This makes wide binaries useful for calibrating parallaxes for this purpose were only available for (relatively) small stellar models as well as spectroscopic and astrometric surveys. samples of nearby and bright stars. Gaia DR2 (Gaia Collaboration

arXiv:2101.05282v3 [astro-ph.SR] 2 Feb 2021 At angular separations greater than about one arcsecond, wide et al. 2018) dramatically expanded the sample of stars with well- binaries are easily resolvable as two point sources. Distinguishing measured parallaxes and proper motions, enabling the construction of physically bound binary stars from chance alignments (“optical dou- unprecedentedly pure and extensive wide binary samples. El-Badry bles”) has been a long-standing challenge for binary star astronomy. & Rix(2018, hereafter ER18) searched Gaia DR2 for pairs of stars Indeed, the ﬁrst systematic binary star catalog was constructed under within 200 pc of the Sun with parallaxes and proper motions consis- the assumption that all close pairs were chance alignments (Her- tent with being gravitationally bound, and projected separations of schel 1782), an assumption that was only shown to be incorrect two up to 50,000 AU (0.24 pc). Their catalog prioritized purity over com- decades later (Herschel 1803). pleteness, and thus imposed relatively strict cuts on astrometric and For bright binaries at close angular separations, chance alignments photometric quality and signal-to-noise ratio (SNR). This resulted in can be excluded probabilistically. However, the contamination rate a catalog of ∼ 55, 000 binaries with an estimated contamination rate from chance alignments increases at wider separations and fainter of ∼ 0.1%. magnitudes. Inclusion of proper motion data can aid the selection of Using the same basic strategy but less stringent cuts on astrometric SNR, Tian et al.(2020) extended the ER18 binary search to larger distances (푑 < 4 kpc) and wider separations (푠 < 1 pc). This pro- ★ E-mail: [email protected] duced a substantially larger sample of ∼ 800, 000 binary candidates,

© 2021 The Authors 2 El-Badry et al. but with a higher contamination rate: chance-alignments dominate and parallax_over_error > 5 their catalog at 푠 & 20, 000 au, though higher-purity subsamples can and parallax_error < 2 be selected by imposing stricter cuts on astrometric SNR. Their cat- and phot_g_mean_mag is not null alog contained 325,000 binaries with 푠 < 20, 000 au; it is expected to be reasonably pure in this separation regime. The query returns a total of 푁 = 64, 407, 853 sources, corresponding Another binary catalog was produced by Hartman & Lépine to 푁 (푁 + 1)/2 ≈ 2 × 1015 possible pairs. Of these, we consider as (2020), who combined Gaia DR2 astrometry with a catalog of high- initial binary candidates all pairs that satisfy the following: proper motion stars not contained in Gaia DR2. In order to reduce • Projected separation less than 1 parsec: the angular separation contamination from chance alignments, they limited their search to between the two stars, 휃, must satisfy binaries with proper motions larger than 40 mas yr−1; this translates 휃 휛 to a distance limit of order 200 pc for typical stars in the Galactic ≤ 206.265 × 1 , (1) disk, but to a larger search volume for stars on halo-like orbits with arcsec mas large tangential velocities. Their primary binary catalog contains where 휛1 is the parallax of the star with the brighter 퐺 magnitude. ∼ 100, 000 binary candidates, with projected separations as large as The maximum search radius of 1 pc (corresponding to an orbital 10 pc. period of ∼ 108 years) is chosen because a vanishingly small number In this paper, we use Gaia eDR3 data (Gaia Collaboration et al. of bound binaries are expected to exist at separations wider than this, 2020a) to further expand the sample of known wide binaries. Com- where the Galactic tidal field becomes comparable to the gravita- pared to DR2, eDR3 astrometry is based on a ∼1.5 × longer time tional attraction of the two stars (e.g. Binney & Tremaine 2008). The baseline. This yields significant improvements in parallax and espe- separation beyond which the Galactic tidal field dominates a binary’s cially proper motion uncertainties. For example, the median uncer- internal acceleration is called the Jacobi radius. In the Solar neigh- parallax pmra tainties in and (one-dimenstional proper motion) at borhood, it is given by 푟퐽 ≈ 1.35 pc × (푀tot/푀 ), where 푀tot is 퐺 = 18 have respectively improved from 0.165 mas to 0.120 mas, the total mass of the binary (Jiang & Tremaine 2010). At separations −1 −1 and from 0.280 mas yr to 0.123 mas yr (Lindegren et al. 2018, slightly below 푟퐽 , binaries are efficiently disrupted by gravitational 2020b). Improvements at the bright end are more significant, due to perturbations from objects such as other stars and molecular clouds better handling of systematics; i.e., at 퐺 = 13, the median parallax (e.g. Weinberg et al. 1987). uncertainty decreased from 0.029 mas to 0.015 mas. This improved • Parallaxes consistent within 3 (or 6) sigma: the parallaxes of astrometric precision allows us to distinguish bound binaries from the two components, 휛1 and 휛2, must satisfy chance alignments to larger distances and wider separations than √︃ was possible with DR2. Our approach is a compromise between the | − | 2 + 2 휛1 휛2 < 푏 휎휛,1 휎휛,2, (2) strategies adopted by Tian et al.(2020) and ER18. Like Tian et al. (2020), we search out to wide separations and relatively low astro- where 휎휛,푖 is the parallax uncertainty of the 푖-th component, and metric SNR, so that at wide separations and faint magnitudes, the 푏 = 3 for pairs with 휃 > 4 arcsec, or 푏 = 6 for pairs with 휃 < 4 full catalog is dominated by chance alignments. However, we also arcsec. The less stringent cut at 휃 < 4 arcsec is adopted because empirically estimate and assign each binary a probability that it is a the chance alignment rate there is low and parallax uncertainties are chance alignment, making it straightforward to select pure subsets of significantly underestimated at close angular separations (Section5). the catalog. • Proper motions consistent with a Keplerian orbit: The two stars The remainder of the paper is organized as follows. In Section2, we in a wide binary will have proper motions that are similar, but, due describe how our binary candidate sample is selected and cleaned. to orbital motion, not identical. We require Section3 describes how we quantify the contamination rate from Δ휇 ≤ Δ휇 + 2휎 , (3) chance alignments and estimate the probability that each binary can- orbit Δ휇 didate is bound. Section4 details basic properties of the catalog and where Δ휇 is the observed scalar proper motion difference, 휎Δ휇 its a cross-match with the LAMOST survey. In Section5, we use the uncertainty, and Δ휇orbit the maximum proper motion difference ex- binary sample to validate the Gaia eDR3 parallax uncertainties. We pected due to orbital motion. The first two quantities are calculated summarize and discuss our results in Section6. The public catalog as is described in Section7. Details about the calculation of chance- 1/2 h ∗ ∗ 2 2i alignment probabilities are provided in the Appendix. Δ휇 = (휇훼,1 − 휇훼,2) + (휇 훿,1 − 휇 훿,2) , (4) 1/2 1 2 2 2 2 2 2 휎 = 휎 ∗ + 휎 ∗ Δ휇 + 휎 + 휎 Δ휇 , (5) Δ휇 휇 휇 훼 휇훿,1 휇훿,2 훿 Δ휇 훼,1 훼,2 2 SAMPLE SELECTION Δ 2 = ( ∗ − ∗ )2 Δ 2 = ( − )2 where 휇훼 휇훼,1 휇훼,2 and 휇 훿 휇 훿,1 휇 훿,2 . Here To reduce contamination from chance alignments, we limit our sam- ∗ 휇훼,푖 ≡ 휇훼,푖 cos 훿푖, 훼 and 훿 denote right ascension and declina- ple to pairs in which both components have moderately precise as- tion, and 휇훼 and 휇 훿, the proper motion in the right ascension and trometry. We retrieved from the Gaia archive all sources with paral- declination directions. Following ER18, we take laxes greater than 1 mas (corresponding to a nominal distance limit 3/2 −1/2 of 1 kpc, which in reality is blurred due to parallax errors), fractional −1 휛 휃 Δ휇orbit = 0.44 mas yr × , (6) parallax uncertainties less than 20%, absolute parallax uncertainties mas arcsec − less than 2 mas, and non-missing 퐺 band magnitudes. This was which is the maximum proper motion difference expected for a cir- achieved with the following ADQL query: cular orbit of total mass 5푀 ; it corresponds to a projected physical velocity difference select* from gaiaedr3.gaia_source 푠 −1/2 Δ푉 = 2.1 km s−1 × , (7) where parallax > 1 orb 1000 au

MNRAS 000,1–26 (2021) A million binaries from Gaia 3

−1 where 푠 = 1000 au × (휃/arcsec)(휛/mas) is the projected separa- initial candidate pairs tion. This quantity is of course not equal to the full 3D separation, or to the semimajor axis, 푎. For randomly oriented orbits with a realistic eccentricity distribution, 푠 and 푎 usually agree within a factor of 2 (see ER18, their Appendix B). For our purposes, it is generally an acceptable approximation to assume 푎 ∼ 푠. We note that the cut on proper motion difference removes from the sample a significant fraction of unresolved hierarchical triples and higher-order multiples (see Section 4.4). We apply the cuts on projected separation, parallax difference, and proper motion difference to all possible pairs. The projected sepa- N = 112, 473, 599 10 ration cut reduces the list of possible binaries to 10 ; additionally (1% plotted) requiring consistent parallaxes and proper motions reduces it to 108 after cleaning initial candidate pairs. A random sample of 1% of these is plotted in the top panel of Figure1, and their projected separation distribution is shown in gray in Figure2. A large majority of initial candidate pairs are chance alignments, not genuine binaries. This is evident both from the fact that many candidate pairs are in the Galactic bulge, LMC, and SMC, and from the fact that their separation distribution peaks at the widest separations (Figure2), where true binaries are rare. Stars in the bulge, LMC, and SMC would ideally be excluded by the requirement of 휛 > 1 and 휛/휎휛 > 5 in the initial query, but a sig- N = 1, 817, 594 nificant fraction of sources in crowded fields have spurious parallaxes (e.g. Fabricius et al. 2020). Most of these spurious background pairs high bound probability can be excluded by imposing astrometric quality cuts – for example, we find that adding the requirement of astrometric_sigma5d_max < 1 to our initial query reduces the number of initial candidates by a factor of 10 while only removing a minority of genuine binaries1. We opted against applying such cuts because they do remove some real binaries, and we find that chance alignments with spurious parallaxes can be efficiently filtered out by the cleaning described in Section 2.1. In addition to chance alignments of background sources with spurious parallaxes, our initial selection also efficiently selects members of star clusters and moving groups, about 100 of which can be seen N = 1, 256, 400 in the top panel of Figure1. These are not spurious, in the sense that they really do contain many pairs of stars within our search volume high bound probability; d < 200 pc that are close in phase space and in some cases mutually bound (e.g. Oh et al. 2017). However, most of them are not binaries and will become unbound when the clusters dissolve.

2.1 Cleaning clusters, background pairs, and triples We clean the list of initial binary candidates in several passes. First, beginning with all the sources returned by our initial ADQL query, we count for each source the number of phase-space neighbors that N = 119, 195 are brighter than 퐺 = 18 and consistent with the size and velocity dispersion of a typical cluster. We define neighboring sources as those that satisfy the following: Figure 1. Sky distribution of binary candidates in Galactic coordinates. Top panel shows a random subset of all pairs with consistent parallaxes and proper • Projected separation less than 5 pc; i.e., 휃 ≤ 17.19 arcmin × motions. The vast majority of these are not genuine binaries. Many clusters ( / ) 휛 mas . are visible, as well as the Magellanic clouds and inner Galaxy. These are −1 • Proper motions within 5 km s ; this translates to a proper mo- beyond the nominal 1 kpc search limit but enter the dataset due to spurious −1 tion difference Δ휇 ≤ 1.05 mas yr × (휛/mas), with a 2휎 tolerance. parallaxes. 2nd panel shows the sky distribution after clusters, moving groups, and resolved triples have been filtered (Section 2.1). This removes the majority of spurious background pairs. 3rd panel shows the pairs with R < 0.1 (corresponding approximately to 90% bound probability; see Section 3.2). 1 astrometric_sigma5d_max is the longest principal axis in the 5- Some structure remains, primarily tracing dust. Bottom panel shows sources dimensional error ellipsoid, in mas. A large value indicates that at least within 200 pc, which are distributed almost uniformly on the sky. one of the astrometric parameters is poorly determined in the 5-parameter solution. More information on this and other Gaia flags can be found in the Gaia eDR3 data model.

MNRAS 000,1–26 (2021) 4 El-Badry et al.

The binary candidate sample after removal of resolved triples and initial candidate pairs 107 suspected cluster members is shown in the 2nd panel of Figure1 after cleaning and with the black histogram in Figure2. This is the catalog pub- 6 10 high bound probability lished along with this work; it contains both high-confidence binaries 105 and pairs that are very likely chance alignments. The contamination rate from chance alignments increases rapidly with separation 4 10 (Section3), while the true binary separation distribution decreases N 103 monotonically over the range of separations to which we are sensitive (e.g., ER18). The separation at which the number of binary 2 10 candidates per dex of separation begins to increase (log(푠/au) ≈ 4.5 101 for the black histogram in Figure2) thus marks the separation at which chance alignments begin to dominate the full sample. That is, 0 10 when considering the full catalog, a majority of binary candidates 1 2 3 4 5 with 푠 & 30, 000 au are chance alignments. It is, however, possible log(s/au) to select subsets of the catalog that are free of chance alignments out to much wider separations; see Section3. We define the “primary” and “secondary” components, denoted Figure 2. Separation distribution of binary candidates at three stages of “1” and “2”, as the component with the brighter and fainter 퐺 mag- the selection procedure, corresponding to the top three panels of Figure1. nitude, respectively. Both components are on the main sequence in a Most initial candidates (gray) are chance alignments, which dominate at majority of binaries; in these cases, the “primary” is generally also 푠 & 3, 000 au. Cleaning resolved triples, clusters, and moving groups removes the more massive component. For binaries containing white dwarfs, a large fraction of chance alignments (black), but they still dominate at 푠 & the secondary will often be more massive than the primary. 30, 000 au. The separation distribution of binaries with high bound probability (& 90%; red; see Section 3.2) falls off steeply at wide separations.

3 CHANCE ALIGNMENTS • Parallaxes consistent within 2 sigma; i.e., Δ휛 ≤ √︃ 2 휎2 + 휎2 . The contamination rate from chance alignments depends on a variety 휛,1 휛,2 of factors, including angular separation, parallax and proper motion, We remove from our binary candidate list all pairs in which either their respective uncertainties, and the local source density. We use component has more than 30 neighbors as defined above. Only 6.5 two complementary approaches to constrain the chance alignment million of the 64 million sources in the search sample have more rate for different subsets of the catalog. than 30 neighbors, and inspection reveals that a majority of these are First, we repeat our binary selection procedure on the GeDR3mock not in the search volume at all, but rather are spurious sources in the catalog produced by Rybizki et al.(2020). This catalog is built on Galactic bulge, LMC and SMC. Removing candidates containing a realization of the Besançon model of the Milky Way (Robin et al. these sources shrinks the candidate list from 112,473,599 pairs to 2003) produced with Galaxia (Sharma et al. 2011). It contains no 2,881,543 and removes most of the obvious structure seen in the top binaries, so by construction all binary candidates selected from the panel of Figure1. mock catalog are chance alignments. It does, however, contain a re- Next, we remove all overlapping pairs. That is, if either component alistic population of open clusters, a variety of stellar populations, of a binary candidate is a member of another binary candidate, we an approximation of the Gaia eDR3 selection function, and realistic remove both pairs. This removes genuine resolved triples, which astrometric uncertainties. We repeat the full binary selection process are efficiently identified by our initial search (e.g. Perpinyà-Vallès described in Section2 on the mock catalog, including filtering of et al. 2019) and are not rare (Tokovinin 2014). It also removes some clusters and resolved triples. We remove pairs with angular separa- additional chance alignments. This cut shrinks the candidate sample tions 휃 < 0.5 arcsec by hand, because the Gaia eDR3 sensitivity from 2,881,543 to 1,918,362. drops precipitously at closer separations (Fabricius et al. 2020), and Finally, we search for members of small clusters or moving groups this is not accounted for in the mock catalog. The separation distri- not removed in the first pass. Adopting the phase-space coordinates bution of candidates selected from the mock catalog is shown in the of the brighter component of each pair as representing the pair, we left panel of Figure3. count the number of neighboring pairs for each candidate, defined Second, we produce an empirical chance alignment sample based using the same three criteria used when counting neighboring sources on the actual Gaia eDR3 catalog, following the method introduced by (without any magnitude cut). We reject all candidates that have more Lépine & Bongiorno(2007). Prior to selecting potential binary com- than 1 neighboring pair, shrinking the sample from 1,918,362 to panions to each star, we artificially shift it from its true position by 1,817,594. ≈ 0.5 degrees, increasing its reported RA by 0.5 sec(훿) degrees. We It is important to note that some real binaries will be removed then repeat the binary search, treating each star’s shifted coordinates during the filtering of resolved triples, clusters, and moving groups. as its true coordinates when searching for possible companions. This In regions of high stellar density, a distant tertiary candidate that is process avoids selecting real binaries, since stars are shifted away really a chance alignment can cause a genuine binary to be rejected from their true companions, but preserves chance alignment statis- as a triple. Similarly, some bound binaries do exist within clusters, tics, because the source density within our 1 kpc search volume does and these will all be rejected. An upper limit of ∼ 15% can be set on not vary much on 0.5 degree scales. Copying and shifting the cat- the fraction of true binaries lost during cleaning by comparing the alog effectively doubles the number of possible chance alignments, gray and black histograms in Figure2 at close separations; this is an increasing the total number of pairs from 푁 (푁 + 1)/2 to 푁 (푁 + 1), upper limit because some of the pairs removed at close separations are so we retain members of the shifted chance alignment catalog with genuinely members of resolved triples, moving groups, and clusters. 50% probability. We again remove pairs with 휃 < 0.5 arcsec. The

MNRAS 000,1–26 (2021) A million binaries from Gaia 5

binary candidates binary candidates 5 5 10 chance alignments 10 chance alignments (GeDR3mock) (shifted catalog) 4 4 10 10

3 3 10 10 N

2 2 10 10

1 1 10 10

0 0 10 10 1 2 3 4 5 1 2 3 4 5 log(s/au) log(s/au)

Figure 3. Two methods for estimating the contamination rate from chance alignments. Left panel compares the separation distribution of binary candidates (black) to that of candidates selected from the eDR3 mock catalog (Rybizki et al. 2020; cyan), which contains no true binaries. Right panel compares the same binary candidates (black) to the separation distribution of candidates when stars are artificially shifted 0.5 degrees from their true positions when selecting candidate companions (orange; by construction, these are not true binaries). Both methods show that chance alignments dominate the full catalog at wide separations (푠 & 30, 000 au, where the separation distribution begins to increase), but true binaries dominate at closer separations. separation distribution of chance alignments produced in this way is that remains pure beyond 푠 ∼ 105 au (∼ 0.5 pc) without also dra- shown in the right panel of Figure3. matically reducing the sample size. We make the catalog of shifted The two methods predict similar chance alignment rates. At wide chance alignments publicly available in order to facilitate estimation separations (log(푠/au) & 4.5 for the full catalog), the number of of the chance alignment rate in various subsamples of the catalog. chance alignments is similar to the number of pairs in the binary candidates catalog. In this regime, most candidates are chance alignments. Not surprisingly, chance alignments begin to dominate at the 3.2 Estimating chance alignment probabilities separation where the binary candidate separation distribution begins As illustrated in Figure4, chance alignments and true binaries are to increase. The number of chance alignments per logarithmic sepa- found in different, but overlapping, regions of parameter space. We ration interval increases steeply with separation: there are 100 times estimate the probability that a particular binary candidate is bound more chance alignments with 5 < log(푠/au) < 5.1 than there are by comparing, at its location in parameter space, the local density with 4 < log(푠/au) < 4.1. This is a consequence of the larger avail- of binary candidates and that of chance alignments from the shifted able area for background stars to be found in at wider separations. At catalog. The process is described in detail in AppendixA. The “den- fixed distance, the area in which chance alignments can appear scales sities” are evaluated in a seven-dimensional space using a Gaussian as ∼ 2휋푠 d푠. At wide separations, the chance alignment distributions kernel density estimate (KDE). The dimensions are (1) angular sepa- thus scale as d푁/d log 푠 ∼ 푠2. This scaling does not hold exactly, in ration, (2) distance, (3) parallax difference uncertainty, (4) local sky part due to small-scale clustering and in part due to the less strin- density, (5) tangential velocity, (6) parallax difference over error, and gent parallax consistency required at small angular separations, but it (7) proper motion difference over error. We re-scale these quantities provides a good approximation to the chance alignment rate at wide to all have similar, order-unity dynamic range before fitting KDEs to separations (log(푠/au) 4). & both the binary candidate and the chance alignment distributions. We use the “shifted” chance alignment catalog in the rest of our We denote the KDE-estimated density of chance alignments at analysis, because we find that it reproduces the separation distribution a point 푥ì in the 7-dimensional parameter space as N (푥ì), of binary candidates in the large-separation limit somewhat more chance align and that of binary candidates as Ncandidates (푥ì). The latter quantity reliably than the mock catalog. is expected to be the sum of the chance-alignment and true binary densities. We then calculate the ratio of these two quantities,

3.1 Chance alignment rate for subsets of the catalog R (푥ì) = Nchance align (푥ì) /Ncandidates (푥ì) . (8) Although chance alignments dominate the full sample at 푠 & This ratio approximately represents the probability that a binary 30, 000 au, it is possible to select subsets of the catalog that are candidate at position 푥ì is a chance alignment, so selecting only can- free of chance alignments out to wider separations. This is illustrated didates with small R is an eﬃcient method for eliminating chance in Figure4, which shows the separation distributions of various alignments. R is not strictly a probability – for example, it is not subsets of the binary candidate and the shifted chance alignment strictly less than one (Figure5) – but it is a serviceable approxima- catalogs. As expected, the chance alignment rate at ﬁxed separation tion for one. We calculate R values for all members of the binary is lower for samples with high galactic latitude, bright component candidate and chance alignment catalogs. stars, small fractional parallax errors, large space velocities, or large Figure5 (left) shows the distribution of R values for both catalogs. proper motions. Nevertheless, it is challenging to select any subset There is a narrow population of binary candidates with R near zero;

MNRAS 000,1–26 (2021) 6 El-Badry et al.

5 all b > 30 deg G < 15 binary candidates 10 | | 2 chance alignments Ngood = 1309715 Ngood = 448123 Ngood = 156078 4 10

3 10 N 2 10

1 10

0 10 1 1 5 ( /σ )2 > 20; |b| > 30 deg; v > 100 km s− v > 250 km s− 10 (∆µ ∆µ )/σ < 1 − orbit ∆µ Ngood = 195924 Ngood = 42763 Ngood = 1303 4 10

3 10 N 2 10

1 10

0 10 1 5 WDMS WDWD µtot > 200 mas yr− ; 10 (∆µ ∆µ )/σ < 1 − orbit ∆µ Ngood = 16944 Ngood = 1458 Ngood = 3750 4 10

3 10 N 2 10

1 10

0 10 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 log(s/au) log(s/au) log(s/au)

Figure 4. Contamination rate from chance alignments in various subsets of the catalog. Black histograms show binary candidates. Orange histograms show a catalog of chance alignments constructed by shifting stars 0.5 degrees when selecting candidate binaries. It contains no real binaries, but it has similar chance alignment statistics to the true binary candidates catalog. For the set of cuts illustrated in each panel, the vertical dashed line shows the widest separation at which there are more than 2× the number of binary candidates as chance alignments. 푁good is the number of binary candidates in that subset closer than this separation. Panels show the full catalog, high galactic latitudes, binaries with both components brighter than 퐺 = 15, precise parallaxes at high Galactic latitudes, high tangential velocities (e.g. halo-like orbits), WD+MS and WD+WD binaries, and high-proper motion pairs. these are objects that have a high probability of being bound. There is excluded from the high-confidence sample than chance alignments a second population of candidates with R ∼ 1; these objects are likely will be included in it. chance alignments. The separation distributions of binary candidates We validate the use of R as a proxy for the chance-alignment prob- with R below several thresholds are shown in the right panel of Fig- ability in Figure6, which compares the Gaia DR2 radial velocities ure5. As expected, the vast majority of binary candidates with close (Sartoretti et al. 2018; Katz et al. 2019) of the two components of separations (log(푠/au) . 4) have low R values, indicating a high candidates in which both stars have measured RVs. We only plot −1 bound probability. At wider separations, the separation distribution candidates in which both components have 휎RV < 10 km s and of high-probability binaries falls off precipitously. We emphasize that the separation is wider than 50, 000 au, where the full catalog is this drop-off is steeper than that of the separation distribution of all dominated by chance alignments. One expects the RVs of the two binaries, because for a low R threshold, more true binaries will be components to be similar for genuine wide binaries. For chance alignments, the RVs of the two components should be drawn from a broad

MNRAS 000,1–26 (2021) A million binaries from Gaia 7

6 all 10 binary candidates 5 chance alignments 10 < 0.5 R < 0.1 4 R 5 10 < 0.01 10 R 3 10 N

4 2 10 10

1 10 3 10 0 10 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1 2 3 4 5 (~x) log(s/au) R

Figure 5. Left: ratio of the local density of chance alignments to binary candidates (Equation8), for all pairs in the binary candidate sample (black) and the shifted chance alignment sample (orange). Likely chance alignments have R ∼ 1; objects with high probability of being bound have R ∼ 0. Right: separation distribution of binary candidates below diﬀerent R thresholds. The chance-alignment rate increases at wide separations, so the separation distribution of low-R candidates falls oﬀ more steeply for progressively smaller R thresholds.

distribution with width comparable to the local velocity dispersion Classification 푁candidates 푁R<0.1 Description of the Galactic disk, and thus they will generally be inconsistent. For MSMS 1,412,903 877,416 both MS the full sample, there are indeed plenty of pairs with obviously in- WDMS 22,563 16,156 one WD, one MS consistent RVs, as chance alignments dominate at wide separations, WDWD 1,565 1,390 both WD even for bright pairs (e.g. Figure4). Red points in Figure6 show MS?? 378,877 360,180 one MS, one no colors binaries with R < 0.1. As expected, these pairs all have RVs close to WD?? 646 547 one WD, one no colors the one-to-one line and are likely all bound. We stress that RVs are ???? 1,040 711 both no colors R not used in creating the catalog or calculating values, so this re- Total 1,817,594 1,256,400 sult bolsters our confidence in the chance alignment ratios calculated from the shifted catalog. Table 1. Contents of the binary candidate catalog. All stars with measured We note that only bright stars (퐺 . 13.5) had RVs published 퐺BP − 퐺RP colors are classified as “MS” or “WD” depending whether they in Gaia DR2 (with no new RVs added in eDR3), so the fraction fall above or below the dashed lines in Figure8. 푁candidates is the number of of chance alignments among binaries where both components have candidate binaries with separations up to 1 pc (black histogram in Figure2); RVs is lower than in the full sample at the same separation. RVs for 푁R<0.1 is the number with high bound probability (red histogram in Figure2). fainter stars are compared in Section 4.5, where we take RVs from the LAMOST survey. In the rest of the paper, we define the “high bound probability” or magnitude diagram (CMD): defining 푀퐺 = 퐺 + 5 log (휛/100), we “high confidence” subset of the catalog as the subset with R < 0.1; classify as WDs objects with 푀퐺 > 3.25 (퐺BP − 퐺RP) + 9.625; this corresponds approximately to > 90% probability of being bound. all other stars with measured 퐺BP − 퐺RP colors are classified as This does not mean that 10% of the pairs in this subset are chance MS stars. Under the ansatz that the two components have the same alignments: most candidates in it have R 0.1 (Figure5). Interpret- distance, we use the (usually more precise) parallax of the primary, ing R as the probability that a given pair is a chance alignment, we (the brighter star) for both components when calculating 푀퐺. estimate that 4,600 of the 1.26 million candidates with R < 0.1 are The adopted WD/MS boundary in the CMD is shown with a dashed chance alignments (0.4%). For R < 0.01, the same fraction is 870 line in Figures7 and8. The boundary is not entirely unambiguous out of 1.15 million (0.08%). We make the full candidate catalog and – particularly for the WD+MS binaries, there are a few objects near R values available, including pairs that are likely chance alignments. the boundary that may be misclassified – but a majority of objects do fall clearly on the WD or MS sequences. We note that the “MS” classification serves only to exclude WDs. The CMDs in Figure7 show that while most non-WD stars are indeed on the main sequence, 4 BASIC PROPERTIES OF THE CATALOG the “MS” class also includes some giants, subgiants, pre-main se- Basic properties of the binary catalog are shown in Figure7 and listed quence stars, and brown dwarfs. The number of binary candidates in Table1. Only the high-confidence pairs ( R < 0.1) are shown. and high-confidence binaries in each class is summarized in Table1. Following ER18, we classify stars as white dwarfs (WDs) or main- The data acquisition window for BP/RP spectra is 2.1 × 3.5 arcsec sequence (MS) based on their position in the Gaia color-absolute wide, preventing colors from being measured for most close pairs

MNRAS 000,1–26 (2021) 8 El-Badry et al.

s > 50, 000 au secondaries in the WD+WD sample are 19.1 and 19.8. For WD+MS binaries, the same values are 15.5 for the primaries (in most cases, 100 all candidates the MS star) and 19.4 for the secondaries. The median parallax high bound probability of the full high-conﬁdence sample is 2.05 mas (1/휛 ≈ 485 pc); the 75 median distances for WD+WD and WD+MS binaries are 148 and 212 parsecs. Because of their closer distances, the WD+WD and WD+MS 50 binaries are distributed roughly uniformly on the sky. The MS+MS

] sample, which extends to larger distances, bears clear imprint of 1

− 25 the stratification of the Galactic disk. Most of the binaries in the s sample are part of the kinematic “disk” population, with a median m 0 −1 −1 −1 k tangential velocity 푣⊥ = 4.74 km s ×(휇 /mas yr )(휛/mas) ≈ [ tot −1 2 35 km s . There is also evidence of a kinematic “halo” population −1 V 25 with 푣⊥ & 200 km s that contains a few thousand binaries. R The WD+MS and WD+WD binaries are shown separately on the 50 CMD in Figure8. On top of the WD+WD binaries, we plot WD cooling tracks for carbon-oxygen cores with hydrogen atmospheres 75 (Holberg & Bergeron 2006; Kowalski & Saumon 2006; Tremblay et al. 2011; Bergeron et al. 2011; Bédard et al. 2020). WDs cool as 100 they age, moving from the upper left to the lower right of the CMD. 100 75 50 25 0 25 50 75 100 Cooling ages are indicated with triangular symbols along the tracks, 1 which mark intervals of 1 Gyr. The faintest WDs in the catalog have RV1 [km s− ] implied cooling ages of about 10 Gyr. 31 WD+WD binaries have one component that falls below the 1.2푀 cooling track; i.e., with a photometrically-implied mass 푀 > 1.2푀 . For WD+MS binaries, Figure 6. Gaia RVS velocities of the components of binary candidates wider than 50,000 au. These are not used in constructing the catalog or calculating 75 WDs fall below the 1.2푀 cooling track. Hydrogen-atmosphere bound probabilities, but are useful for validation. Black points show all can- cooling tracks are not appropriate for WDs with non-DA spectral −1 didate binaries for which both components have 휎RV < 10 km s ; red points types, so spectroscopic classification must be obtained before masses show the subset of these with R < 0.1, corresponding approximately to a and ages of individual WDs can be inferred with high fidelity. 90% bound probability. These all fall close to the one-to-one line, suggesting The catalog also contains about 10,000 high-confidence binaries that they are indeed bound. in which the primary is a giant (about half these giants are in the red clump), including about 130 giant-giant binaries. These are all quite bright (both components have 퐺 < 11). They, along with the (Arenou et al. 2018). The majority of sources with a comparably WD+MS binaries and the ≈ 13, 000 binaries in which one component bright companion within 2 arcsec thus do not have measured 퐺 − BP is a subgiant, can serve as useful calibrators for stellar ages. 퐺RP colors and cannot be classified as WD or MS stars based on Gaia data alone. Components lacking a color measurement are denoted Massive stars are not well-represented in the catalog; there are “??” in Table1 and in the catalog. About 30% of all high-confidence 351 high-confidence primaries with 퐺BP − 퐺RP < 1 and 푀퐺 < binaries have one component with unknown color; 0.05% lack colors 0. This cut corresponds roughly to 푀 & 3푀 , though extinction for both components. In 98% of cases where only one component complicates the mapping between 푀퐺 and mass. 75% of the high- has a color, it is the brighter component. In many cases where no confidence MS+MS binaries have primaries with 푀퐺 between 9.5 Gaia color is available, colors from other surveys (e.g. Pan-STARRS) and 3.8, corresponding approximately to 0.4 < 푀1/푀 < 1.3. For should be sufficient to distinguish WD and MS components. secondaries, the 75% range is 푀퐺 = 11.3 − 6.4, corresponding to The catalog contains 8.8 × 105 high-confidence MS+MS bina- 0.2 < 푀2/푀 < 0.8. The catalog also contains about 80 binaries in ries, more than 16,000 high-confidence WD+MS binaries, and 1,390 which one component is likely a brown dwarf. We identify these on high-confidence WD+WD binaries. Angular separations range from the CMD as objects with 푀퐺 > 16.5 and 퐺 − 퐺RP > 1.3; they are 0.2 arcsec to one degree. The peak of the angular separation dis- all within 80 pc of the Sun and therefore are also found in the binary tribution is at 1.2 arcsec. This is simply a result of the Gaia eDR3 catalog produced by Gaia Collaboration et al.(2020b). angular resolution, since the intrinsic separation distribution falls off A small but noticeable fraction of sources, particularly secon- monotonically with increasing separation over all separations that daries, are scattered below the main sequence in the CMD, between are well-represented in the catalog (e.g. Duchêne & Kraus 2013). the WDs and MS stars (bottom right panel of Figure7). The majority There are 271 pairs with separations between 0.2 and 0.4 arcsec, of these sources are separated from a brighter companion by only including 24 below 0.3 arcsec. Fabricius et al.(2020, their Figure a few arcsec; the most likely explanation for their anomalous CMD 7) found some indication that the Gaia eDR3 catalog may contain position is thus that they are MS stars with contaminated BP/RP pho- spurious duplicated sources at separations below 0.4 arcsec (that is, tometry. The majority of such sources can be filtered out using cuts on single sources that were erroneously classified as two sources), so bp_rp_excess_factor and phot_bp/rp_n_blended_transits it is possible that a small fraction of the closest pairs in the catalog (e.g. Riello et al. 2020), but we refrain from employing such cuts are spurious. We do not, however, find any increase in the angular since they also remove a significant fraction of sources with ac- separation distribution at close separations, as might be expected to ceptable photometry and astrometry. Some sources below the main arise from a population of duplicated sources. sequence may also be stars with spurious parallaxes or biased colors, The median magnitude of high-confidence primaries is 퐺 = 15.2, and a few are likely real astrophysical sources, primarily cataclysmic and that of secondaries is 퐺 = 17.7. Most WDs in the catalog variables and detached but unresolved WD+MS binaries (e.g. Abra- are significantly fainter: the median magnitudes of primaries and hams et al. 2020; Belokurov et al. 2020).

MNRAS 000,1–26 (2021) A million binaries from Gaia 9

5 5 5 10 all=1,256,400 10 all 10 MSMS=877,416 MS+MS 4 WDMS=16,156 4 4 10 WDWD=1,390 10 WD+MS 10 WD+WD 3 3 3 10 10 10

N 2 2 2 10 10 10

1 1 1 10 10 10

0 0 0 10 10 10 1 2 3 4 5 0 1 2 3 5 10 15 20 log(s/au) log(θ/arcsec) G1 [mag]

5 10 5 10 4 10 4 4 10 10 3 3 3 10 10 10

N 2 2 2 10 10 10

1 1 1 10 10 10

0 0 0 10 10 10 0 5 10 15 0.00 0.25 0.50 0.75 1.00 50 0 50 G G [mag] 1/ [kpc] b [galactic latitude; deg] 1 − 2

5 0 0 10 primaries secondaries

4 10 5 5 3

10 G G N M M 2 10 10 10

1 10

0 15 15 10 0 200 400 600 0 2 4 0 2 4 1 G G G G v [km s− ] BP − RP BP − RP

Figure 7. Basic properties of the high-bound probability binary sample: physical and angular separation, apparent magnitude of the primary, magnitude difference, distance, galactic latitude, plane-of-the-sky velocity, and color-magnitude diagrams. We designate stars as white dwarfs (WDs) or main sequence (MS; including giants) based on whether they fall below or above the dashed line in the bottom right panels. Black, blue, and yellow histograms show properties of MS+MS, WD+MS, and WD+WD binaries. We only classify binaries as MS+MS, WD+MS, or WD+WD if both components have a 퐺BP − 퐺RP color. A majority of binaries with separation 휃 < 2 arcsec do not have Gaia colors for a least one component and are therefore not classified; this accounts for most of the difference between the black and gray histograms (see also Table1).

MNRAS 000,1–26 (2021) 10 El-Badry et al.

WD + MS binaries WD + WD binaries MS stars primaries 0.0 8 WDs secondaries 2.5 M = 0.6M ¯ 10 M = 0.9M 5.0 ¯ M = 1.2M ¯

G 7.5 G 12 M M 10.0 14 12.5

15.0 16

17.5 1 0 1 2 3 4 5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 G G G G BP − RP BP − RP

Figure 8. Color–absolute magnitude diagrams of high-conﬁdence WD+MS and WD+WD binaries. Left: black points show MS components, which are usually the primaries; WD components are shown in red. Right: primary and secondary WDs are shown in black and red. We overplot cooling models for hydrogen-atmosphere WDs with masses between 0.6 and 1.2 푀 . Symbols mark 1 Gyr intervals of cooling age, with the ﬁrst (leftmost) symbol at 1 Gyr.

4.1 Twin binaries and contrast sensitivity 9 chance alignments An excess population of equal-brightness (and presumably, equal- 6

] mass) “twin” binaries is also found in the catalog. Its existence is

g 3 most obvious in the distribution of magnitude diﬀerence, Δ퐺, as a a function of separation, which is shown in Figure9 and compared to m [ 0 chance alignments. Unlike the chance alignment catalogs constructed G 3 from the mock catalog and shifted catalogs (Section3), these are

∆ selected in the same way as true binaries, but with the requirement 6 that the parallaxes and proper motions of the two components be inconsistent. For easier visualization of the distribution of magnitude 9 diﬀerence near Δ퐺 = 0, the sign of Δ퐺 is randomized. The top panel 9 binaries illustrates the separation-dependent contrast sensitivity of the Gaia 6 eDR3 catalog: at close angular separations, sources with signiﬁcantly

] brighter companions are outshone. This leads to a contrast limit of g 3 a Δ퐺 ≈ 4 mag at 휃 = 1 arcsec and Δ퐺 ≈ 7 mag at 휃 = 2 arcsec. The

m contrast limit at a given separation is not “sharp”, but is manifest [ 0 as a smooth drop in sensitivity with increasing Δ퐺 (e.g. Brandeker G 3 & Cataldi 2019).2 The contrast sensitivity is signiﬁcantly improved

∆ in the binary catalog produced in this work compared to the one 6 produced by ER18: that work required both components to have 9 relatively uncontaminated BP/RP colors and thus contained basically 1 2 3 4 5 no binaries closer than 2 arcsec. θ [arcsec] The bottom panel of Figure9 shows true binary candidates, in which the two components do have consistent parallaxes and proper motions. Unlike with the chance alignments, here there is an narrow excess population with Δ퐺 ≈ 0. The extent and provenance of this Figure 9. Magnitude diﬀerence between the two stars in chance alignments population was studied by El-Badry et al.(2019); here we simply note (top) and bound wide binaries (bottom) as a function of angular separation. The sign of Δ퐺 is randomized. Chance alignments are selected following the that it is also clearly apparent in our catalog. Because the twin excess same procedure used for real binaries, but with the requirement that the two is most prominent at close physical separations, it is somewhat more stars’ parallaxes be inconsistent rather than consistent. The chance alignments obvious in our catalog, which extends to closer angular and physical illustrate the contrast sensitivity of the Gaia eDR3 separations. catalog: at close separations, pairs with large magnitude diﬀerence are not detected. In the bottom panel, a narrow excess population of binaries with Δ퐺 ≈ 0 is visible, which is absent in the top panel. This highlights the excess population of equal-mass “twin” binaries. 2 At very close separations (휃 . 0.7 arcsec; not shown in Figure9), the Gaia eDR3 catalog contains only equal-brightness pairs (Lindegren et al. 2020b, their Figure 6). The contrast sensitivity is relatively smooth at 휃 > 1 arcsec.

MNRAS 000,1–26 (2021) A million binaries from Gaia 11

4.2 Comparison to other catalogs over 5000 < 푠/au < 50, 000. The separation distribution from Hart- man & Lépine(2020) has a similar logarithmic slope to the one from Figure 10 compares the distributions of projected physical separation, ER18 in this separation range and likely tracks the intrinsic separation angular separation, and distance of the catalog produced here to other distribution out to wider separations. At the very widest separations wide binary catalogs in the literature: represented in that catalog (5+ pc, exceeding the local Jacobi radius • The ER18 catalog (magenta) was produced from Gaia DR2 and corresponding to an orbital period of about a Gyr), it is unlikely using a similar strategy to this work, but it was limited to binaries that pairs are actually bound. This may reflect the fact that the search within 200 pc (휛 > 5 mas) and used more stringent cuts on both strategy employed by Hartman & Lépine(2020) is sensitive to any photometric and astrometric quality and SNR. In order to maintain pairs that are closer in phase space than chance alignments from the high purity, it only contains binary candidates with 푠 < 50, 000 au. shifted catalog, without explicit consideration of the expected orbital • Tian et al.(2020) expanded the ER18 search strategy to a larger velocities. That is, their search does not distinguish between bound volume (4 kpc) and used less stringent quality cuts, while still using binaries, moving groups, or stellar streams. In terms of absolute numbers, the catalog represents a factor of Gaia DR2 data. Unlike this work, which uses a cut of 휛/휎휛 > 5 ∼ 4 increase in the number of high-confidence binaries over the one for both components, they required 휛/휎휛 > 20 for the primary and from Tian et al.(2020). To our knowledge, it is the largest published 휛/휎휛 > 2 for the secondary. This results in somewhat different contamination properties and completeness. They also searched out catalog of high-confidence binaries of any type. The sample could to 푠 = 1 pc. In addition to to their full catalog of all candidates, Tian likely be expanded by a further factor of a few by loosening the et al.(2020) published 3 smaller catalogs with high purity, which are distance and parallax uncertainty limits, or dropping the parallax cut not shown in Figure 10. entirely, while focusing on close angular separations (e.g. Dhital et al. • Hartman & Lépine(2020) did not use a strict distance cut, 2015). However, the cuts we use in this paper provide a reasonable but limited their search to high-proper-motion pairs with 휇 > compromise between sample size, purity, and data quality. 40 mas yr−1. This preferentially selects nearby stars, since proper A wide binary catalog based on Gaia eDR3 was also produced by motion is inversely proportional to distance at fixed transverse veloc- Gaia Collaboration et al.(2020b), which contains pairs within 100 −1 pc. It is not shown in Figure 10, but we find that within 100 pc, it is ity. For a “typical” tangential velocity of 푣⊥ = 35 km s , their proper motion cut corresponds to 푑 < 185 pc. However, stars on halo-like or- almost identical to ours. bits with larger tangential velocities are included to larger distances; −1 −1 e.g., a binary with 푣⊥ = 200 km s will have 휇 > 40 mas yr out 4.3 Space density to a distance of 1.05 kpc. Rather than employing strict cuts on parallax and proper motion consistency, Hartman & Lépine(2020) used Figure 11 compares the separation distributions of binaries in differ- empirical estimates of the chance alignment rate as a function of ent (cumulative) distance bins. We scale these by the effective stellar position and proper motion difference from a shifted catalog (sim- volume corresponding to the distance cut, i.e., by a factor propor- ilar to our approach in Section3) to distinguish true binaries from tional to the total number of stars expected in each distance sample. chance alignments. Their approach has the advantage of not requir- Were it not for the finite scale height of the Galactic disk, this factor 3 ing specific cuts in parallax or proper motion difference, which are would simply be the search volume 푉 = 4휋푑max/3, where 푑max is always somewhat arbitrary. A disadvantage is that it does not account the distance limit. We approximate the total stellar density within for the heteroskedasticity of parallax and proper motion uncertain- our search volume as a plane-parallel exponential distribution with ties – i.e., the chance alignment probability is higher for pairs with the Sun at the midplane and a scale height ℎ푧 = 300 pc (Jurić et al. large astrometric uncertainties, and this is not accounted for in their 2008). We then define an effective volume 푉˜ , which represents the analysis. number of stars in a sphere of radius 푑max divided by the stellar density at the disk midplane: One difference between the catalog produced in this work and ∫ 푑max ˜ −푧/ℎ푧 2 2 the other catalogs is obvious in Figure 10: our sample extends to 푉 = 2휋 푒 푑max − 푧 푑푧 (9) smaller angular separations, and thus also physical separations. This 0 h i 3 −푑max/ℎ푧 2 −푑max/ℎ푧 2 is partly a result of the improved angular resolution of Gaia eDR3 = 2휋 2ℎ푧 푒 − 1 + 2ℎ푧 푑max푒 + ℎ푧 푑max . (e.g., Fabricius et al. 2020) but is primarily due to a change in search (10) strategy. ER18 and Tian et al.(2020) required both components of 3 candidate binaries to have 퐺BP −퐺RP colors and to pass photometric As expected, this expression asymptotes to 4휋푑max/3 in the limit of quality cuts related to the bp_rp_excess_factor reported in Gaia 푑max ℎ푧. In Figure 11, the separation distributions for each value DR2 (Evans et al. 2018). This set a soft resolution limit of ∼ 2 arcsec, of 푑max are divided by the appropriate value of 푉˜ . with a wider effective limit for pairs with large brightness contrast. Dashed vertical lines in Figure 11 mark a separation of 푠res limit = We do not require colors or employ a photometric quality cut in (2 arcsec)×푑max. At separations 푠 < 푠res limit, incompleteness due to this work; this adds an additional ∼400,000 binary candidates to the the Gaia eDR3 angular resolution starts to become severe. The figure sample that would be excluded if we did (Table1). shows that for 푑max . 200 pc, incompleteness is due primarily to the Both the catalog produced in this work and the one from Tian et al. angular resolution limit: at 푠 > 푠res limit, the separation distributions (2020) become dominated by chance alignments at 푠 & 30, 000 au. in different distance bins overlap. However, for 푑max = 500 pc or 1000 The “high bound probability” subset of our catalog does not, but pc, the catalog contains fewer binaries per effective volume than at its separation distribution falls off steeply at wide separations. This closer distances, even at 푠 > 푠res limit. This reflects the fact that at decline at wide separations is steeper than that of the intrinsic sep- sufficiently large distances, some binaries will have components that aration distribution, since a decreasing fraction of binaries at wide are too faint to pass the parallax_over_error > 5 limit, or to be separations can be identified as bound with high confidence. This can detected at all. At 푑 = 200 pc, parallax_over_error > 5 implies be seen in comparing the red separation distribution to the magenta 휎휛 < 1 mas. This is satisfied by most sources with 퐺 . 20.5 one from ER18, which tracks the intrinsic separation distribution (Lindegren et al. 2020b), corresponding to 푀퐺 = 14.0, near the

MNRAS 000,1–26 (2021) 12 El-Badry et al.

5 high bound probability 5 5 Tian + 20 10 10 10 all Hartman + 20 ER18 4 4 4 10 10 10

3 3 3 10 10 10 N

2 2 2 10 10 10

1 1 1 10 10 10

0 0 0 10 10 10 0 1 2 3 4 5 6 0 1 2 3 4 0 1 2 3 4 log(s/au) log(θ/arcsec) 1/ [kpc]

Figure 10. Comparison of the catalog produced in this work (black and red histograms) to the wide binary catalogs published by ER18, Tian et al.(2020), and Hartman & Lépine(2020), all of which are based on Gaia DR2 data. This catalog expands the one produced by Tian et al.(2020) by a factor of ∼ 4. The full catalog (black) has lower purity than the El-Badry & Rix(2018) and Hartman & Lépine(2020) catalogs at wide separations. The subset with high bound probability (red) maintains high purity at wide separations, but does not purport to trace the intrinsic separation distribution there.

] lognormal separation distribution for solar-type binaries (Raghavan 1 0.005 −

x d < 15 pc et al. 2010). We caution that eﬀects of astrometric acceleration also e

d d < 25 pc become important in this regime (Section 6.3), potentially leading to

3 0.004

− d < 50 pc spurious parallaxes and preventing pairs from being recognized as c

p d < 100 pc [ binaries by our search.

e 0.003 d < 200 pc Integrating over separation, the distributions in Figure 11 imply a n a l d < 500 pc total space density of (0.006±0.001) wide binaries with 푠 > 30 au per p

d d < 1000 pc i 0.002 cubic parsec in the solar neighborhood. For context, the space density

m of all unresolved Gaia eDR3 sources in the solar neighborhood is t −3 a 0.07 pc (Gaia Collaboration et al. 2020b), about 10 times higher.

s 0.001

g When all members of multiple systems are counted individually, the o

l −3

d total stellar space density in the solar neighborhood is about 0.10 pc / 0.000 n (e.g. Winters et al. 2020). d 1 2 3 4 5 log(s/au) 4.4 Orbital velocities from proper motion differences Precise parallaxes and proper motions make it possible to estimate Figure 11. Midplane number density of binaries in different cumulative dis- orbital velocities (projected onto the plane of the sky) from the proper tance bins. Vertical lines mark 휃 = 2 arcsec at the distance limit of each motion difference between the two stars. The plane-of-the sky veloc- bin, roughly the angular separation at which incompleteness due to blend- ity difference Δ푉 can be calculated as ing becomes significant. At all distances, the turnover at close separations is driven primarily by the resolution limit, but the 15 and 25 pc samples show a relatively flat separation distribution below a few hundred au. Beyond 200 Δ휇 휛 −1 Δ푉 = 4.74 km s−1 × . (11) pc, there is significant incompleteness due to stars being too faint to pass our mas yr−1 mas fractional parallax error threshold – i.e., the calculated number densities are lower than in the nearby samples even at wide separations, where blending Here Δ휇 is the scalar proper motion difference (Equation4), and 휛 should not be important. is the parallax of the binary, for which we take the parallax of the brighter component. The corresponding uncertainty is √︄ bottom of the main sequence (e.g. Figure8). That is, the sample 2 휎2 −1 (Δ휇) 2 Δ휇 is expected to be almost complete for 푑 = 200 pc, except for 휎Δ푉 = 4.74 km s 휎 + , (12) max 휛4 휛 휛2 crowding/blending effects at close separations. On the other hand, −1 at 푑 = 1 kpc, parallax_over_error > 5 implies 휎휛 < 0.2 mas, with 휎Δ휇 calculated from Equation5, Δ휇 and 휎Δ휇 in mas yr , and 퐺 . 18.8, and 푀퐺 . 8.8, meaning that most of the lower main 휛 and 휎휛 in mas. We implicitly assume here that the two stars have sequence will be excluded. the same parallax. Equation 12 is almost always dominated by the first All of the separation distributions in Figure 11 increase to- term under the radical; i.e., parallax errors dominate over proper mo- ward smaller separations at 푠 > 푠res limit, but the distributions for tion errors. The median value of 휎Δ푉 for all high-confidence binaries −1 −1 푑max = 15 pc and 푑max = 25 pc are similar and do appear to flat- in the catalog is 0.33 km s ; 195,601 have 휎Δ푉 < 0.1 km s . ten above the resolution limit, at 푠 ∼ 30 au. For a typical binary in Figure 12 explores the Δ푉 values of binary candidates in the the catalog with total mass 1 푀 , this corresponds to a period of catalog, and the effects of the Δ푉 cuts we employ on its purity and order 200 years, which is indeed near the peak of the approximately completeness. The upper left panel shows all pairs from our initial

MNRAS 000,1–26 (2021) A million binaries from Gaia 13

1 1 σ∆V < 0.1 km s− ; ∆V < 5 km s− all candidates candidates; high bound probability chance alignments triples

] 0 1 10 − s m k

[ binaries V

∆ -1 10

∆Vorb

1 1 chance alingments; σ∆V < 0.1 km s− ; ∆V < 5 km s− chance alignments (all) chance alignments; high bound probability

] 0 1 10 − s m k [ V

∆ -1 10

2 3 4 5 2 3 4 5 2 3 4 5 10 10 10 10 10 10 10 10 10 10 10 10 s/au s/au s/au

−1 Figure 12. Upper left: plane-of-the sky velocity difference between the components of wide binary candidates selected with Δ푉 < 5 km s and 휎Δ푉 < −1 0.1 km s . Dashed red line shows the maximum expected velocity difference for a binary with total mass 5푀 (Equation7). Binary candidates in our catalog (top center) are required to be consistent, within 2 휎, with falling below this line, but pairs with large proper motion uncertainties can scatter well above it. The clouds of points at large separation and Δ푉 are chance alignments; bottom panels show chance alignments from the shifted catalogs for the same selection. Points at large separation and Δ푉 are excluded by the requirement of high bound probability (R < 0.1; upper right). In the upper left panel, there is a population of triples and higher-order multiples above the dashed line, with no corresponding population in the shifted catalog (lower left). These have increased Δ푉 due to the gravitational effects of an unresolved close companion. They are generally gravitationally bound, but are excluded from our primary catalog. query that have 푠 < 1 pc, consistent parallaxes, Δ푉 < 5 km s−1, least one component (Figure 13), suggesting that these components −1 and 휎Δ푉 < 0.1. Note that the cut of Δ푉 < 5 km s is generally are unresolved binaries (e.g. Belokurov et al. 2020). less strict than the one we adopt in constructing our primary catalog A large fraction of these unresolved triples and higher-order mul- (Equation3), which is equivalent to Δ푉 < Δ푉orb +2휎Δ푉 , with Δ푉orb tiples are excluded from our actual catalog by the requirement of −1 given by Equation7. Δ푉 < Δ푉orb + 2휎Δ푉 rather than e.g., Δ푉 < 5 km s ; this should In the upper left panel, there is a clear ridgeline of binaries with be kept in mind when using the catalog for applications involving Δ푉 ∼ 푠−1/2, as expected from Kepler’s laws. This population largely higher-order multiplicity. We do not use a constant Δ푉 cut (e.g. −1 falls below Δ푉orb (dashed red line) because most of the binaries in Δ푉 < 5 km s ) in constructing the full catalog because this would the catalog have total masses less than 5푀 . There is not, however, result in a much higher contamination rate from chance alignments. −1 a sharp drop-off at Δ푉 > Δ푉orb. The population with Δ푉 > Δ푉orb For 5, 000 < 푠/au < 10, 000, 22% of pairs with Δ푉 < 5 km s and −1 likely consists primarily of triples and higher-order multiples, in 휎Δ푉 < 0.1 km s are excluded from the catalog by Equation3. If which the plane-of-the-sky velocity of one component is affected by we assume that large Δ푉 values are due mainly to subsystems, this a close, unresolved companion (e.g. Clarke 2020; Belokurov et al. implies that about 12% of wide binary components in this separation 2020). It is also possible that some of these pairs are moving groups range have an unresolved subsystem that imparts a large enough pho- that are not actually bound but remain close in phase space (e.g. tocenter perturbation to significantly increase Δ푉. This corresponds Pittordis & Sutherland 2019; Coronado et al. 2020); however, we find to a subsystem separation range of ∼ 1 − 100 au, since photocenter that the population exists even among binaries on halo-like orbits, wobbles will average out over the 34-month Gaia eDR3 baseline favoring multiplicity as the primary explanation for it. We also find at closer separations, and the subsystem-induced Δ푉 will be small 3 that most pairs with Δ푉 > Δ푉orb have unusually large ruwe for at at wider separations. About 25% of wide binary components have a subsystem in this separation range (Tokovinin & Smekhov 2002;

3 ruwe, the re-normalised unit-weight error, is a measure of astrometric of-ﬁt indicators with magnitude and color. Values above about 1.4 indicate goodness-of-ﬁt that corrects for global trends in the other reported goodness- potential problems.

MNRAS 000,1–26 (2021) 14 El-Badry et al.

max ruwe alog with Gaia eDR3 using the CDS Xmatch service,4 which uses 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Gaia proper motions to propagate source positions to epoch J2000. We matched each LAMOST observation to the nearest Gaia source within 1 arcsec. For sources with more than one LAMOST observation, we retained only the observation with the highest 푔−band SNR. This left us with LAMOST data for 4,306,131 sources, which we

] 0 then matched to our catalog using Gaia source ids. 1 10

− This yielded 91,477 binaries in which at least one component has s a LAMOST spectrum. This sample will be useful for a variety of m

k applications, such as studying the dependence of the binary fraction [ on metallicity (e.g. El-Badry & Rix 2019; Hwang et al. 2020). Here, V both ∆ we focus on a subset of the cross-match: those binaries in which -1 10 components have a LAMOST spectrum, and the angular separation is at least 3 arcseconds. The latter cut is to avoid cases where both ∆Vorb stars fall inside a single fiber, leading to potentially biased stellar 2 3 4 5 10 10 10 10 parameters and abundances (e.g. El-Badry et al. 2018). The RVs of s/au these binaries can be used to verify whether most binary candidates are bound. This test is similar to that shown in Figure6 with Gaia RVs, but LAMOST spectra extend to fainter magnitudes than RVs from Gaia, which are currently only available at 퐺 . 13. Figure 13. Binary candidates from the upper left panel of Figure 12, now Figure 14 compares the LAMOSTs RVs and metallicities (i.e., colored by the maximum ruwe of the two components. Most pairs with 5 [Fe/H]) of the components of these binaries. These are generally Δ푉 > Δ푉orb and 푠 10 au (those labeled “triples” in Figure 12) have ruwe & 1.4 for at least one component, in most cases because that component expected to be consistent for genuine wide binaries. The RVs for is an unresolved binary. At 푠 . 500 au, most pairs have large ruwe due to most binaries do indeed fall close to the one-to-one line, but there centroiding errors at close angular separations (see also Figure 18). are some outliers: 91 of the 2840 binaries in the sample have RVs that are more than 3-sigma discrepant. A potential worry is that these pairs are not binaries at all, but chance alignments. To assess whether Tokovinin et al. 2010), so it is quite plausible that the large-Δ푉 pairs this is likely to be the case, we plot the separation distribution of the are mostly unresolved triples and higher-order multiples. discrepant pairs in the bottom right panel of Figure 14. This shows The clouds of points at large separation and Δ푉 in Figure 12 that the separation distribution of pairs with large RV differences are chance alignments. These are primarily pairs with large 휎Δ푉 ; is similar to that of all pairs. Chance alignments are much more otherwise they would be excluded by the requirement of Δ푉 < common at wide separations (Figure4), so if chance alignments were Δ푉orb + 2휎Δ푉 . The bottom panels show the distribution of chance the root of the discrepant RVs, one would expect these pairs to be alignments from the shifted catalog (Section3). These are distributed clustered at large separations. The primary reason for the discrepant similarly to the large-separation cloud among binary candidates. No RVs is likely again higher-order multiplicity. This is likely to affect chance-alignment cloud is visible in the binary candidates with high the LAMOST RVs more than it does those from Gaia (Figure6), bound probability (upper right), but a few chance alignments do because the LAMOST RVs and their uncertainties are based on a scatter into the bound binary cloud, mostly at Δ푉 > Δ푉orb (lower single epoch. Unresolved short-period binaries are usually filtered right). out of the Gaia RV sample we consider, because 휎RV for that sample We note that the interpretation of Δ푉 calculated from Equation 11 is calculated from the epoch-to-epoch RV dispersion, and we required −1 as a physical velocity difference between the two components of a 휎RV < 10 km s for both components. binary breaks down at large angular separations, where projection The lower left panel of Figure 14 shows the distribution of effects become important (e.g. El-Badry 2019). Indeed, two stars in uncertainty-normalized RV difference between the two components an ultra-wide binary can have identical space velocities but substan- √︃ of binaries; i.e., ΔRV/휎 = (RV − RV ) / 휎2 + 휎2 . The tially different plane-of-the-sky proper motions. The magnitude of ΔRV 1 2 RV,1 RV,2 −1 the apparent proper motion difference depends primarily on angular median 휎ΔRV for this sample is 7 km s , which is larger than the −1 separation and is therefore largest for nearby binaries, which are also typical ∼ 1 km s orbital velocity for these binaries, so we expect the the binaries with the smallest 휎Δ푉 . For the sample shown in the width of the main distribution to be dominated by measurement un- upper left panel of Figure 12, projection effects become important certainties, with the tails dominated by higher-order multiples. If the beyond about 20,000 au, which is – perhaps not coincidentally – the reported 휎RV values are accurate, ΔRV/휎ΔRV should be distributed separation beyond which the trend of Δ푉 ∼ 푠−1/2 appears to flatten. as a Gaussian with 휎 ≈ 1, with some outliers at higher velocity These projections effects can be corrected if the RVs of at least one difference. The distribution is indeed approximately Gaussian, but it component are known (El-Badry 2019). is narrower than 휎 = 1; the bulk of the distribution is better approxi- mated by 휎 ≈ 0.7. This suggests that the LAMOST RV uncertainties are typically overestimated by ∼ 30%. 4.5 Cross-match with LAMOST The distribution of uncertainty-normalized [Fe/H] differences tells a different story. It is not well-described by a single Gaussian, but We cross-matched the binary catalog with the LAMOST survey (Cui has a narrow component with 휎 . 1, and broad, asymmetric tails. et al. 2012, DR6 v2), the currently most extensive spectroscopic sur- This suggests that 휎 values are considerably underestimated vey providing stellar parameters and abundances. We began with the [Fe/H] LAMOST low-resolution “A, F, G and K Star” catalog, which contains atmospheric parameters and metallicities for 5,773,552 spectra (including some duplicate observations). We cross-matched the cat- 4 http://cdsxmatch.u-strasbg.fr/

MNRAS 000,1–26 (2021) A million binaries from Gaia 15

150 0.5 N = 2840 500 primaries 100 secondaries ] ] 400 1 x − 50 e 0.0 s d [

m 300 2 ] k 0 [ H 2 / 0.5 e V 50 200 F R [

100 100 1.0 150 0 150 100 50 0 50 100 150 1.0 0.5 0.0 0.5 5 10 15 20 1 RV1 [km s− ] [Fe/H]1 [dex] G [mag]

0.4 0.6 σ = 1 σ = 1 ∆RV/σ∆RV > 3 | | ) σ = 0.7 ] σ = 2 all H / ) 0.5 e 0.3 2 V F

[ 10 R ∆ ∆ 0.4 σ σ / / ]

V 0.2 0.3 H /

R 1 e 10 ∆ F [ ( 0.2 p

∆ 0.1 (

0.1 p 0 10 0.0 0.0 4 2 0 2 4 5 0 5 2 3 4 5 ∆RV/σ∆RV ∆[Fe/H]/σ∆[Fe/H] log(s/au)

Figure 14. High-confidence binaries in which both components were observed by the LAMOST survey. We only include pairs with angular separation 휃 > 3 arcsec to avoid blending. Top left and center: comparison of the RVs and metallicities ([Fe/H]) of the primary (푥-axis) and secondary (푦-axis). In most binaries, the two components have very similar RV and [Fe/H], as expected. There is evidence of a systematic bias toward lower [Fe/H] for fainter targets. Upper right: 퐺 magnitudes of both components. This sample includes fainter stars than the Gaia RVS sample. Bottom left and center: distributions of uncertainty-normalized RV and [Fe/H] difference, compared to Gaussians with the listed 휎. These suggest that the LAMOST RV uncertainties are overestimated by ∼ 30% on average, while [Fe/H] uncertainties are underestimated and subject to temperature systematics. Bottom right: separation distributions. Pairs in which RVs are not consistent within 3휎 are shown in red. If these pairs were chance-alignments, one would expect them to be concentrated at the widest separations. They are not, suggesting that they are either higher-order multiples with the RV of one component biased by an unresolved close companion, or pairs in which one component has a catastrophically wrong RV. for a significant fraction of the catalog. We find (not shown in the et al.(2020) to assess the reliability of abundances derived by the figure) that the distribution becomes narrower and more Gaussian GALAH survey. when we only consider binaries in which the two components have similar magnitude and effective temperature. This suggests that the larger-than-expected metallicity differences are due primarily to tem- 5 CALIBRATING GAIA DR3 PARALLAX UNCERTAINTIES perature systematics in the abundance pipeline. This is particularly evident in the upper middle panel of Figure 14, which shows that Because the two stars in a wide binary have very nearly the same the secondaries (which have lower 푇eff) systematically have lower re- distance, our catalog provides a straightforward method of validating ported [Fe/H]. It is of course possible that some binaries really have the Gaia eDR3 parallax uncertainties. We do this by calculating the inconsistent surface abundances, but work with higher-quality spec- uncertainty-normalized parallax difference between the two compo- √︃ tra than those which underlie the LAMOST metallicities suggests Δ / = ( − ) / 2 + 2 nents of each binary, 휛 휎Δ휛 휛1 휛2 휎휛,1 휎휛,2, just such abundance anomalies are rare (Hawkins et al. 2020). as we did for the LAMOST RVs. In the limit of accurate parallax uncertainties (and small differences in the true distance to the two Many of the binaries in our catalog were also observed by other components), this quantity should be distributed as a Gaussian with spectroscopic surveys. We defer analysis of these data to future work, 휎 = 1. If the reported parallax uncertainties are underestimated, one but comment that the type of analysis shown in Figure 14 will be expects a wider distribution, and possibly deviations from Gaussian- useful in calibrating the abundances derived by surveys (and their ity. uncertainties). For example, an earlier version of our binary catalog Figure 15 shows distributions of Δ휛/휎Δ휛 for binaries at a range constructed from Gaia DR2 was recently fruitfully used by Buder of angular separations and magnitudes. To isolate underestimated

MNRAS 000,1–26 (2021) 16 El-Badry et al.

11 < G < 12; 0 < θ[00] < 2 11 < G < 12; 2 < θ[00] < 4 11 < G < 12; 4 < θ[00] < 10

0.4 σ = 1.70 0.4 σ = 1.48 0.4 σ = 1.28 ) N = 351 σ = 1 N = 551 N = 490